Developer uses GPT-6 to jailbreak a Kindle over Wi-Fi and turn it into a photo frame

APPSO · wechat · 2026-10-18

Developer Niraj Rajgor had an AI agent jailbreak a 10th-gen Kindle Basic over Wi-Fi — no USB — using the open-source SpiderCat exploit from the KindleModding community, then turned the device into a looping photo frame.

Model trials: GPT-5.6 Sol (Extra High reasoning) got the helper app installed but failed on the Kindle filesystem; Anthropic models tripped safety guards after two prompts and got downgraded to Opus 4.8; GPT-6 Astra (High reasoning) finished the job. OpenAI's report calls it the first model rated Critical under its Preparedness Framework for cybersecurity, with reduced CoT monitorability.

Key workflow insight: the agent couldn't see the Kindle screen, so early debugging relied on Rajgor photographing the display and uploading it. Adding remote screenshots and automatic log retrieval let the agent map the filesystem on its own.

Physical-world blind spot: all automated tests passed on the Mac, yet the real e-ink screen showed ghosting from the previous image — screenshots capture the system's frame, not the physical refresh of e-ink particles. The piece cites Anthropic's October 2026 research on AI models self-replicating via IoT devices as the flip side of this capability.

Original post →

More from coding & agent

coding & agent channel →